Perfect Reconstruction AM-FM Image Models

For the first time, we present an AM-FM image model that, in addition to being remarkably consistent with human visual perception, also provides perfect reconstruction of the image and is thus suitable for synthesis as well as analysis applications. We employ a non-separable coiflet-based wavelet filterbank with channels that are both orientation selective and jointly localized. The analysis responses define the AM-FM image components, which we demodulate analytically using a new, high-quality 2-D phase unwrapping algorithm coupled with a spline-based phase model. The amplitude and frequency modulations obtained with this approach correspond remarkably well with human visual perception of the salient image structures, suggesting that this invertible model could form the basis for a general theory of image processing in the modulation domain.